Qwen3.8 Next on older hardware: DDR4 + 7900 XTX hits 45–50 t/s
10/03/2026 — 10/03, 12:12·1 sources·1 reports
Story overview
On October 3, 2026, a user on Reddit's LocalLLaMA forum posted results from running Qwen3.8 Next on an older machine built around DDR4 memory and a 7900 XTX GPU. According to the post, the IQ3_XXS quantized weights — which come in just under 80GB — stabilize at around 45-50 t/s on that setup, with the user noting the figure can run higher at times, depending on mtp, particularly while coding.
For comparison, the same user reported that Llama CPP, after tuning, maxed out at about 22.5 t/s on the same rig. They also said they had seen similar results reported online by users running 12GB and 16GB cards, while DDR5 platforms produced significantly faster numbers.
The post included one caveat: the user advised against using Q2 quantization, and described the quality of the IQ3_XXS run as reliably superior. No further details about the hardware configuration, the exact Qwen3.8 Next variant, or independent verification were provided in the report.
AI-generated from 1 reports · updated 1 hour ago
Latest turnA LocalLLaMA user reports running Qwen3.8 Next at IQ3_XXS (just under 80GB) on DDR4 memory and a 7900 XTX at a steady 45-50 t/s, versus roughly 22.5 t/s at best with a tuned Llama CPP on the same rig. They say users with 12GB and 16GB cards see similar results and DDR5 systems are notably faster, and they advise against the Q2 weights.
Reports on this story headlines open the original
A LocalLLaMA user reports running Qwen3.8 Next at IQ3_XXS (just under 80GB) on DDR4 memory and a 7900 XTX at a steady 45-50 t/s, versus roughly 22.5 t/s at best with a tuned Llama CPP on the same rig. They say users with 12GB and 16GB cards see similar results and DDR5 systems are notably faster, and they advise against the Q2 weights.
Reddit · LocalLLaMAAI score 62
Other stories people are talking about
- 554NVIDIA launches 64GB DGX Spark desktop AI computer at $4,9998 sources
- 539RisingApple tightens macOS Full Disk Access over AI agent risks7 sources
- 400Meta open-sources Muse Gadgets firmware and SDK5 sources
- 234SurgeAmazon Weighs Moving $8 Billion of Nvidia Chips into a Financing Vehicle3 sources
- 227Anthropic launches Claude Frontier Academy with $100M3 sources
- 226SurgeHugging Face open-sources AstaBrief for fast report generation3 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion